1. why pallas? JAX works with pure functions (i.e., same inputs will produce the same outputs). JAX arrays are immutable not flexible or efficient for kernel implementation GEMM steps input matrix →

source: The Ultra-scale Playbook 0. High-level overview Targed on large-scale (like 512 GPUs) training Tradeoff among the following factors memory usage: params, optimizer states, gradients compute ef

1. Memory Composition Model Parameters Intermediate Activations (Forward pass) will be used to calculate gradiants during backward Gradients (Backward pass) Optimizer States 2. Static Memory (Weigh

Given the length of the official JAX tutorial, this note distills the core concepts, providing an quick reference after reading the original tutorial. High-level JAX stack Source: Yi Wang’s linkedin

Q1: are we running out of pretraining data? are we hitting the scaling law wall? I don’t quite buy it. Gemini only use a portion of the video data to train. We spent plenty of time on filtering the r

1. OverallMany inference talks, but more RL talks. RL RL101 3 RL challenges: training collapses, training slow, hardware errors New frameworks / API: Tinker, SkyRL, Slime, SGLang’s Slime-based fr

1. From Rewards to Optimization In RL, an agent interacts with an environment by observing a state , taking an action , and receiving a reward . In the context of LLM, state is the previous tokens, wh

truncated importance sampling (tis) this blog is from feng yao (ucsd phd student)’s work. i added some background and explanations to make it easier to understand. slides: on the rollout-training mis

Speaker: Lily Liu Working at OpenAI Graduated from UC Berkeley in early 2025 vLLM speculative decoding TL 1. Why is LLM generation slow? GPU memory hierarchy. A100 example: SRAM is super fast (19 TB

Speaker: Roger Wang 1. Overview large multi-modal models (LMMs) most SOTA Large Multimodal Models leverage a language model backbone with an encoder for a non-text modality. E.g., LLaVA, Qwen VL, Qwen